SEOF: Reducing Spikes for Efficient SCNN Accelerators with Approximate Computing

Spiking Convolutional Neural Networks (SCNNs), a variant of Spiking Neural Networks (SNNs) based on the architecture of Convolutional Neural Networks (CNNs), have demonstrated superior performance in tasks compared to traditional SNNs. However, this enhanced capability results in higher energy consumption due to increased computation and weight access. Research has extensively demonstrated that SCNN computation is directly related to spike counts. Thus, reducing the number of spikes can lead to more efficient implementations of SCNN inference accelerators. However, arbitrarily reducing spikes often results in an uncontrollable decline in accuracy. In this paper, we present a spike-efficient optimization framework, SEOF, which integrates approximate computing principles and exploits SCNN-specific characteristics to achieve energy savings while maintaining low inference accuracy loss. SEOF incorporates a novel spike-efficient controller for spiking neurons, a spike ratio-oriented objective function designed to produce accelerator-friendly SCNN models, and a spatial mask that lowers energy consumption by reducing weight accesses. Our experimental results indicate that SEOF can achieve 31%-51% computation energy savings by reducing spikes, with an accuracy loss of less than 1%, and a speedup of up to 1.56×. Additionally, SEOF can reduce weight access by up to 54.93%, leading to overall energy savings of up to 50.77%.